Machine Learning Diagnosis of Cleft Speech: Hypernasality Imitation Project

Machine Learning (ML) has been applied to speech and voice analysis for the binary differentiation of Parkinson’s disease from controls (Tsanas et al., 2012; Arora, Babhai-Ravary and Tsanas, 2019) as well as for monitoring progression (Tsanas et al., 2011) and response to treatment (Tsanas et al., 2014). ML has been applied to detect voice and speech changes in Alzheimer disease and other conditions that cause even mild cognitive impairment and neurological conditions (Titze, 2000). Some early work suggests that ML techniques can be used to detect cleft speech characteristics (He et al. 2015; Zhang et al. 2020) . Deep neural networks (DNN) have been trained to detect hypernasality, in limited samples (Mathad et al. 2020) . However, DNNs have limitations including overfitting, which may lead to the development of a model that performs well when used on the dataset on which it was trained, but that fails to detect hypernasality on other datasets. We aim to address these limitations, by adopting a signal processing approach, extracting and analysing over 500 types of signals from audio files, then training ML models to identify unique "signatures" of cleft speech characteristics.

Aims:
1. To develop machine learning (ML) algorithms-
a. to identify and grade acoustic and aerodynamic characteristics unique to cleft speech
b. that can analyse these characteristics irrespective of language/dialect
2. To determine the minimum speech sample (duration and quality) essential for these analyses